Liza

Liza coordinates coding agents through a software delivery workflow with separate authors and reviewers. It's for developers who want to delegate work from specifications through tested, documented code while retaining control over product decisions. Its Go supervisor enforces role boundaries, validation gates and merge authority in code, rather than relying entirely on agents to follow prompts.

The orchestration runs through coding CLIs and Git worktrees, using your existing provider subscriptions and CLI setup. Liza supports Claude Opus through Claude Code and GPT-5 through Codex. It's open source under Apache 2.0; the orchestration software and model access are separate.

You can choose human-led pairing, adversarial pairing for a bounded change, or a full multi-agent workflow for goals that need decomposition. In the full workflow, agents develop specifications, architecture and code plans before implementation, with binding review at each stage. The task graph can adapt when work reveals that the original breakdown needs revision.

A shared behavioral contract asks agents to explain their reasoning, verify claims, disclose uncertainty and stay within scope. Humans can steer between sprints, while agents execute and review within them. Full multi-agent runs require meaningful validation coverage and consume more tokens than the smaller pairing workflows.

Isolated worktrees separate concurrent changes. A live terminal interface shows progress, while shared task records, recorded prompts and agent logs make runs inspectable. Crash recovery, context handoff and a circuit breaker handle interrupted work and repeated failures.